A cortical neural prosthesis for restoring and enhancing memory by theodore w berger robert e hampson dong song anushka goonawardena vasilis z marmarelis and sam a deadwyler

A cortical neural prosthesis for restoring and enhancing memory

This article has been downloaded from IOPscience. Please scroll down to see the full text article.

2011 J. Neural Eng. 8 046017

(http://iopscience.iop.org/1741-2552/8/4/046017)

Download details:

IP Address: 128.125.205.105

The article was downloaded on 08/03/2012 at 23:56

Please note that terms and conditions apply.

View the table of contents for this issue, or go to the journal homepage for more

Home Search Collections Journals About Contact us My IOPscience

IOP PUBLISHING JOURNAL OF NEURAL ENGINEERING

J. Neural Eng. 8 (2011) 046017 (11pp) doi:10.1088/1741-2560/8/4/046017

A cortical neural prosthesis for restoringand enhancing memoryTheodore W Berger1, Robert E Hampson2, Dong Song1,Anushka Goonawardena2, Vasilis Z Marmarelis1 and Sam A Deadwyler2

1 Depatment of Biomedical Engineering, University of Southern California, Los Angeles, CA, USA2 Department of Physiology and Pharmacology, Wake Forest University, Winston-Salem, NC, USA

E-mail: [email protected]

Received 8 November 2010Accepted for publication 11 February 2011Published 15 June 2011Online at stacks.iop.org/JNE/8/046017

AbstractA primary objective in developing a neural prosthesis is to replace neural circuitry in the brainthat no longer functions appropriately. Such a goal requires artificial reconstruction ofneuron-to-neuron connections in a way that can be recognized by the remaining normalcircuitry, and that promotes appropriate interaction. In this study, the application of a speciallydesigned neural prosthesis using a multi-input/multi-output (MIMO) nonlinear model isdemonstrated by using trains of electrical stimulation pulses to substitute for MIMO modelderived ensemble firing patterns. Ensembles of CA3 and CA1 hippocampal neurons, recordedfrom rats performing a delayed-nonmatch-to-sample (DNMS) memory task, exhibitedsuccessful encoding of trial-specific sample lever information in the form of differentspatiotemporal firing patterns. MIMO patterns, identified online and in real-time, wereemployed within a closed-loop behavioral paradigm. Results showed that the model was ableto predict successful performance on the same trial. Also, MIMO model-derived patterns,delivered as electrical stimulation to the same electrodes, improved performance under normaltesting conditions and, more importantly, were capable of recovering performance whendelivered to animals with ensemble hippocampal activity compromised by pharmacologicblockade of synaptic transmission. These integrated experimental-modeling studies show forthe first time that, with sufficient information about the neural coding of memories, a neuralprosthesis capable of real-time diagnosis and manipulation of the encoding process can restoreand even enhance cognitive, mnemonic processes.

1. Introduction

The utility and validation of neural prosthetic deviceshas improved considerably in recent years with severalnew applications providing relief and improvement frompathological circumstances that affect motor control,locomotion, and restoration of movement (Velliste et al 2008,Leuthardt et al 2009, Isa et al 2009, Wang et al 2010a).The primary models for such devices have utilized the well-established assessment of motor cortical neural ensemblefiring related to arm movements and brain–computer interfaces(BCIs) for control of scalp recorded neural events correlatedwith detection of expected sensory events (Daly and Wolpaw2008, Leuthardt et al 2009, Wang et al 2010b). While highlysuccessful, the above approaches have only recently been

applied to cognitive processes within a context that can beshown to facilitate or restore mental activity impaired by(1) events such as lack of attention to or misinterpretationof critical information in normal subjects (Jarosiewicz et al2008, Legenstein et al 2010), or (2) the loss of these capacitiesdue to injury or disease that affect cortical structures involvedin information, encoding, and integration (Moritz et al 2008,Miller et al 2010).

This paper describes the initial demonstration of a corticalneural prosthesis applied to information processing of twosubregions of the hippocampus involved in the formationof long-term memory. These memories were essential forsuccessful performance of a cognitive task requiring encodingand retrieval of information on a trial-by-trial basis. Theneural structure tested by the device, the rodent hippocampus,

1741-2560/11/046017+11$33.00 1 © 2011 IOP Publishing Ltd Printed in the UK

J. Neural Eng. 8 (2011) 046017 T W Berger et al

has been extensively characterized with respect its rolein performance of a delayed-nonmatch-to-sample (DNMS)working memory task (Deadwyler and Hampson 1995, 1997,2006, Hampson et al 1999b, Hampson and Deadwyler 2003,Deadwyler et al 2007) and offline computational assessmentof firing relationships during that performance (Song et al2007a, 2007b). Such characterization provided the substratefor online application of the device to extract critical functionalprocessing from ensembles of hippocampal neurons requiredfor the performance of the task (Deadwyler and Hampson2004, Hampson et al 2008), and to establish when not present,information processing necessary for successful completionof the task (Hampson and Deadwyler 2003, Deadwyler et al2007). Here we demonstrate how this prototype of a corticalprosthesis is capable of (1) monitoring the input pattern tohippocampus during the information encoding phase of thetask, (2) predicting accurately the associated hippocampaloutput pattern and the degree of success related to suchencoding, and (3) delivering stimulation with electrical pulsesduring the same phase of the task in a pattern that conformsto the normal firing of the hippocampal output region onsuccessful trials. The utility of this cortical prosthesis is furtherdemonstrated by the capacity to substitute successful encodingstimulation when hippocampal ensembles are compromisedand cannot generate the necessary codes to perform the tasksuccessfully.

2. Methods

2.1. Subjects and training

Long-Evans rats (Harlan) aged four–six months were usedas subjects (n = 45). All animal protocols and surgicalprocedures are approved by the Wake Forest UniversityInstitutional Animal Care and Use Committee and the NationalInstitute of Health Guide for the Care and Use of LaboratoryAnimals (NIH Publication no 8023). Animals maintainedat 85% of ad libitum body weight were trained to criterionperformance levels in the DNMS task and then implantedbilaterally with two identical array electrodes (Neurolinc, NewYork, NY), each consisting of two rows of eight stainless steelwires (diameter: 20 μ) positioned parallel to the longitudinalaxis of hippocampus (figure 1). The distance between twoadjacent electrodes in the array within a row was 200 μmand between rows was 400 μm to conform to the locationsof the respective CA3 and CA1 cell layers. The electrodearray was lowered to a depth of 3.0–4.0 mm from the corticalsurface with the longer electrode row positioned in the CA3cell layer and shorter row 1.2 mm higher with tips in the CA1layer. Extracellular neuronal spike activity was monitoredduring implantation to maximize placement in the appropriatehippocampal cell layers. Animals were allowed to recoverfrom surgery for at least one week before continuing behavioraltesting (Hampson et al 2008).

2.2. Behavioral apparatus and training procedure

The behavioral testing apparatus for the DNMS task is thesame as reported in prior studies (Deadwyler and Hampson

1997, 2004, Hampson et al 2008) and consists of a 43 ×43 × 50 cm Plexiglas chamber with two retractable levers(left and right) positioned on either side of a water troughon one wall of the chamber (figure 1). A photocell with acue light activated nose-poke (NP) device was mounted inthe center of the opposite wall. The chamber was housedinside a sound-attenuated cubicle. The DNMS task consistedof three phases: (1) sample phase: in which a single lever waspresented randomly in either the left or right position; whenthe animal pressed the lever, the event was classified as sampleresponse (SR), (2) delay phase: of variable duration (1–30 s)in which a nosepoke (NP) into a photocell was required toadvance to the (3) nonmatch phase: in which both levers werepresented and a response on the lever opposite to the SR, i.e.,a nonmatch response (NR), was required for delivery of a dropof water (0.4 ml) in the trough. A response in the nonmatchphase on the lever in the same position as the SR (matchresponse) constituted an ‘error’ with no water delivery and aturning off of the lights in the chamber for 5.0 s. Followingthe reward delivery (or an error) the levers were retracted for10.0 s before the sample lever was presented to begin the nexttrial. Individual performance was assessed as % correct NRswith respect to the total number of trials (100–150) per daily(1–2 h) session.

2.3. Multineuron recording of hippocampal ensembles

Animals were connected by a cable to a 32-channelrecording apparatus via a slip-ring commutator (CristInstruments, Hagerstown, MD) which allowed free movementat all times during behavioral testing. Single neuronaction potentials (spikes) were isolated by time-amplitudewindow discrimination and computer-identified (Plexon Inc.,Dallas, TX, USA) by individual waveform characteristicsusing a multi-neuron acquisition (MAP) processor. Isolatedspike waveforms exhibiting firing rates consistent withCA1 and CA3 principal cells (i.e. 0.5–5.0 Hz baselinefiring rate) which showed stable behavioral correlatesacross sessions (Deadwyler et al 2007, Deadwyler andHampson 2006, Hampson et al 2008) were employedfor experimental manipulations and model development.Hippocampal ensembles used to analyze encoding of DNMPevents consisted of 15–32 single neurons, each recorded froma separate identified electrode location on the bilateral arrays(Song et al 2007a).

2.4. Nonlinear systems identification of ensemble codes

A general, Volterra kernel-based strategy for modeling multi-input/multi-output (MIMO) nonlinear dynamics underlyingspike train-to-spike train transformations between CA3 andCA1 was established to predict output patterns of CA1 firingpattern from input patterns of CA3 neural activity (Bergeret al 2005, 2010 Song et al 2007a, 2009). In this approach, theidentification of spatio-temporal pattern transformations fromthe hippocampal CA3 region to the CA1 region is formulatedas the estimation of an MIMO model that can be decomposedinto a series of multi-input, single-output (MISO) models with

2

J. Neural Eng. 8 (2011) 046017 T W Berger et al

Figure 1. Top, diagram of DNMS task (clockwise from top). Sample phase starts with single lever extended in one position. SR encodingstrength (recorded via array electrodes in CA3 and CA1, center right) predicts delay-sensitive (red arrow) nonmatch response (left).Following random delay, both levers are extended in nonmatch phase. Nonmatch decision requires animal to press lever opposite to SR.Spatiotemporal patterns of ensemble firing are illustrated in color contour maps showing differential firing to left versus right SRs whichoccurred for 3.0–5.0 s following sample lever presentation (SP). Maps depict firing rate of neurons (vertical axis) by time (horizontal axis)for same 5.0 s period. Overall feedback loop for trials is closed by behavioral outcome based on correct or error recall of sample leverposition. Bottom left: performance (mean percentage correct ± SEM) curve illustrates sensitivity of behavior to duration of delay intervalunder control conditions (n = 23). A unilateral cannula located in CA3 is shown next to array to indicate the site of drug infusion to altertask-related firing (figure 4).

physiologically identifiable structure that can be expressed bythe following equations:

w = u(k, x) + a(h, y) + ε(σ ), y ={

0 when w < θ

1 when w � θ.

The variable x represents input spike trains; y representsoutput spike trains. The hidden variable w represents thepre-threshold membrane potential of the output neurons, andis equal to the summation of three components, i.e. post-synaptic potential u caused by input spike trains, the outputspike-triggered after-potential a, and a Gaussian white noiseε with standard deviation σ . The noise term models bothintrinsic noise of the output neuron and the contribution ofunobserved inputs. When w exceeds threshold, θ , an outputspike is generated and a feedback after-potential (a) is triggeredand then added to w. Feedforward kernels k describe thetransformation from x to u. The feedback kernel, h, describes

the transformation from y to a. u can be expressed as a Volterrafunctional series of x, as in

u(t) = k0 +N∑

n=1

Mk∑τ=0

k(n)1 (τ )xn(t − τ)

+N∑

n=1

Mk∑τ1=0

Mk∑τ2=0

k(n)2s (τ1, τ2)xn(t − τ1)xn(t − τ2)

+N∑

n1=1

n1−1∑n2=1

Mk∑τ1=0

Mk∑τ2=0

k(n1,n2)2x (τ1, τ2)xn1(t − τ1)xn2(t − τ2)

+N∑

n=1

Mk∑τ1=0

Mk∑τ2=0

Mk∑τ3=0

k(n)3s (τ1, τ2, τ3)

× xn(t − τ1)xn(t − τ2)xn(t − τ3) + · · ·The zeroth order kernel, k0, is the value of u when the inputis absent. First order kernels, k

(n)1 , describe the linear relation

3

J. Neural Eng. 8 (2011) 046017 T W Berger et al

between the nth input xn and u. Second and third order self-kernels, k

(n)2 , and k

(n)3s , describe the second and third order

nonlinear relation between nth input xn and u, respectively.Second order cross-kernels k

(n1,n2)2x reflect the second order

nonlinear interactions between each unique pair of inputs(xn1 and xn2) as they affect u. N is the number of inputs.Mk denotes the memory length of the feedforward process.The feedback variable a can be expressed as:

a(t) =Mh∑τ=1

h(τ)y(t − τ),

where h is the linear feedback kernel. Mh is the memory lengthof the feedback process. In total, then, the model describeshow temporal patterns of third order (i.e., the effects of triplets)for each input, and second order (i.e., the effects of pairs)for any of two interacting inputs, affect each output, takinginto account differing noise level and output spike-triggeredfeedback (the latter due to circuitry and/or membranebiophysics), and neuron-specific differences in thresholds. Inorder to reduce the number of open parameters to be estimated,both k and h are expanded with orthonormal Laguerre basisfunctions (Marmarelis and Orme 1993, Marmarelis 2004,Marmarelis and Berger 2005). Due to the Gaussian noiseterm and the threshold, this model can be considered a specialcase of the generalized Laguerre–Volterra model (GLVM),which employs a probit link function (Song et al 2007b,2009). All model parameters, i.e., feedforward Volterrakernel k and feedback Volterra kernel h can be estimatedusing an iterative re-weighted least-squares method (Truccoloet al 2005). Noise standard deviation σ and threshold θ

are redundant variables and thus can be indirectly obtainedthrough variable transformation (Song et al 2007b, 2009).Due to the stochastic nature of the system, estimated modelsare validated using an out-of-sample Kolmogorov–Smirnovtest based on the time-rescaling theorem (Brown et al 2002).

Prior studies have shown the success of the application ofMIMO model (Song et al 2007a, 2007b) which was appliedoffline to data from hippocampal ensemble recordings in theDNMS task to determine the temporal relation between spikeoccurrences recorded in CA3 and CA1 (Zanos et al 2008, Songet al 2009). Recent investigations have revealed extensionof this capacity to online control of DNMS performance(Hampson et al 2011). In the latter case it was revealedthat the recordings from the CA3 electrodes constituted theinputs to the model, while the designated outputs werespikes recorded from the CA1 electrodes on the same array(figure 1) which corresponded to the well-establishedmonosynaptic relationship between the two cell populationsvia Schaffer collateral fiber connections (Witter and Amaral2004).

2.5. MIMO generated ensemble stimulation

Stimulation studies were conducted with a custom built 16-channel stimulator (Triangle Bio Systems Inc., Durham, NC)designed to deliver patterns of electrical pulses to CA1electrodes in both (bilateral) hippocampal arrays. Stimulator

output was delivered as bipolar pulses, 1.0 ms duration, 0.1–15 V, 20–100 μA to pairs of CA1 electrodes located adjacentto each other on the array (figure 1). This provided maximalisolation of individual pulse trains from other electrodes inthe same array and restricted stimulation trains to the regionswhere the same firing patterns were obtained from CA1cells recorded on the same electrodes. Stimulation patternsconsisted of 16 channels of pulses delivered bipolar to all CA1electrodes in trains of 1.5–3.0 s duration prior to and duringthe SR, or for control purposes, shifted to 3.0–7.0 s after theSR. Controls for nonspecific effects of stimulation included(1) shuffling of MIMO coefficients to generate stimulationpatterns, or (2) reversing the appropriate pattern for a givenlever position to the pattern for the opposite lever position. Totest of the effectiveness of the electrical stimulation patternssome animals were prepared with minipumps to infuse theglutamatergic blocking agent MK801 into the hippocampusto disrupt MIMO model predicted output patterns in CA1indicated by a loss of identified SR codes. The drug wasinfused chronically (37.5 μg h−1; 1.5 mg ml−1, 0.25 μl h−1)over a two week period in animals that had previously showedfacilitated performance on trials with MIMO stimulationdelivered during a similar chronic infusion of saline vehicle.The location of the infusion cannula is shown in the illustrationof array recordings in figure 1.

3. Results

Prior investigations have demonstrated from online recordingsof hippocampal activity that information encoded during theSR and employed in the nonmatch phase is predictive ofsuccessful or unsuccessful performance based on the durationof the ensuing delay interval (Deadwyler et al 2007, Hampsonet al 2008). These investigations have determined that theretrieval process (NR) is not as critical as the encoding (SR)phase of the task since errors are correlated more with lackof effective hippocampal cell firing during the SR (Simeralet al 2005). DNMS trials are grouped by 5–10 s intervals ofdelay duration, and plotted as mean (± S.E.M.) percentageof correct trials. Animals were trained on trials with delaysof 1–30 s, then subjected on a random infrequent basisto ‘probe’ trials, with delays >30 s, employed to assessaccuracy of the MIMO model and closed loop manipulations.Control performance on ‘probe’ trials typically resulted in near‘chance’ (50% correct) performance and provided a consistentbasis for testing the application of the MIMO model andderived patterns of electrical stimulation. The effectivenessor ‘strength’ of the SR encoding process was evaluated withrespect to whether the trial was rewarded or not as a result ofthe choice of the NR. ‘Strong’ SR codes are defined as MIMOderived firing patterns associated with correct respondingon long delay trials, while ‘weak’ SR codes are definedas MIMO patterns that occur on error (non-rewarded) trialsirrespective of duration of delay. Therefore, performance ofthe task on a given trial is characterized by (1) an interactionbetween ‘strength’ of the SR code (i.e. ‘completeness’ of thespatiotemporal firing pattern exhibited on correct long delaytrials) for sample lever position and (2) duration of the ensuing

4

J. Neural Eng. 8 (2011) 046017 T W Berger et al

Figure 2. Closed loop feedback using MIMO identification of SR encoding strength to control DNMS performance. Top: CA3 and CA1neuronal firing (contour, left) analyzed via MIMO model (center) predicts CA1 firing patterns (contour, right). Strength of SR code in CA1during sample phase predicts behavioral outcome (nonmatch decision) based on prior correlation of firing pattern with success. Closed loop(light blue arrow) reduces (CL-weak codes), or lengthens (CL-strong codes) the delay duration on the same trial to validate MIMO modelprediction of strength of the SR code. Bottom right: mean strong and weak code patterns of probability of neural firing at each neuronposition (array recording site) and time relative to SR. Color contours scaled from blue (<10%) to red (>60%) according to probability ofneural firing on a single trial. Bottom left: mean DNMS performance across animals (n = 15) compares performance on closed loop MIMOmodel-detected strong (CL-strong codes) and weak (CL-weak codes) SR codes to non-closed loop trials (control). Performance onnon-closed loop trials with detected weak codes is displayed as well (weak codes). Performance on trials in which MIMO coefficients werescrambled (scrambled coefficients) is also shown.

delay preceding the nonmatch phase where the encoded SRinformation is retrieved to perform the NR (figure 1).

3.1. MIMO model measure of SR strength of encodingpredicts learned behavior

The method for assessing the relationship between strength ofthe SR code and performance utilized a closed loop paradigm(figure 2) in which CA3 activity occuring 1.5–3.0 s prior tothe SR was used as input to the MIMO model to providea prediction of the output firing pattern of simultaneouslyrecorded CA1 cells (Hampson et al 2011). The strength

of the SR code categorized on the basis of correlationswith prior performance as defined above was utilized toadjust the duration of the delay online during the same trial(figure 2). Depiction of differences in strong versus weakSR codes is shown by the mean firing rate contour mapsin figure 2 (lower right). Because of the extended timeinterval between the SR and NR due to interposed delaysin the task it was possible to adjust the delay duration onthe same trial as a function of the ‘strength’ of the encodedSR, determined online by the MIMO model. The closedloop procedure adjusted delay duration primarily in two ways:(1) on any trial in the session where a weak SR code was

5

J. Neural Eng. 8 (2011) 046017 T W Berger et al

detected the delay was reduced to 10 s, and (2) on trials withdetected strong SR codes, delays were extended to 40, 50,or 60 s. The graph in figure 2 (lower left) shows the twoextreme conditions of closed loop control of performanceby (1) shortening the delay to 10 s for weak SR code trials(CL-weak code) or (2) lengthening the delay to > 30 s whenstrong SR codes were detected (CL-strong codes). It is clearthat performance on strong SR code trials was significantlyelevated (F(1,731) = 25.84, p < 0.001) during normal delays(�30 s) and on extended delay (�40 s) trials, relative toperformance on non-closed loop trials of the same duration(control vs CL-strong code, figure 2). Trials with weak SRcodes adjusted to 10 s delays by the closed loop paradigm wereperformed nearly perfectly by all animals (CL-weak codes,figure 2). However, figure 2 also shows that performance wassignificantly improved on the remaining trials at all delays,(CL-weak code versus control sessions; F(1,731) = 7.96,p < 0.01) due to the fact that weak code trials were eliminatedby the closed loop as a potential source of error on longer delaytrials (weak codes, figure 2).

Verification of the MIMO model predictions is shown fortrials �30 s in which strong or weak SR codes were detectedbut delays not altered by the closed loop procedure (figure 2CL-strong code trials <30 s, weak codes shown by invertedtriangles). It is clear that performance was maximal whenstrong codes were detected (CL-strong codes), but markedlydecreased on all trials >15 s when weak codes were detectedand at risk for error (weak code vs control, F(1, 731) = 15.61,p < 0.001) as shown in figure 2. It is clear that DNMSperformance under normal conditions was a function of twofactors: (1) the strength of the SR code, and (2) the coincidenceof that code strength with duration of the subsequent randomlyassigned delay. Since the delay between SR and NR wasunpredictable, strength of SR encoding on any given trial wasarbitrary and reflected possible ‘anticipation’ of the subsequentdelay duration. However, as shown previously (Hampsonet al 2011), if delays were extended beyond previously trainedlimits (CL-strong code >30 s, figure 2), performance on strongSR code trials also declined in a delay-dependent manner.

3.2. MIMO model stimulation reverses mismatches of SRencoding strength and delay

A unique application of the MIMO model in the closedloop paradigm described above is the ability to substituteelectrical stimulation pulses in patterns that mimicked thefiring of CA1 outputs predicted from the MIMO model,to facilitate memory under circumstances where there weremismatches in the online generated SR code strength forthe duration of the ensuing delay. Such stimulation wasused to enhance memory on trials in which the MIMOmodel detected weak SR codes for trials with delays >10 s(figure 3). Thus substitution of electrical stimulation pulsesat the same electrode loci and in the same temporal pattern asstrong SR codes reversed mismatches between spontaneouslygenerated SR code strength and the duration of subsequent trialdelay (figure 3). This stimulation procedure was capable offacilitating performance in the same manner as MIMO

predicted strong SR codes when delays were extended (30–60 s) in the task (figure 3). Trials with interposed CA1stimulation patterns derived from MIMO generated strongSR codes (stim MIMO model) were significantly increasedversus trials on which no stimulation (no stim) was delivered(F(1,731) = 11.50, p < 0.001). Trials in which stimulationat the same intensity was generated from scrambled MIMOmodel CA1 coefficients (figure 2) did not differ from nostim trials (F(1,731) = 2.12, ns.). Under such circumstancesMIMO model stimulation parameters were applied to thesame electrode locations but in a different spatiotemporalpattern (scrambled coefficients, figure 3). The fact thatstimulation patterns delivered with scrambled coefficients didnot consistently suppress performance as might be expectedrelates to the fact that only some of the scrambled patternsproduced the exact SR codes for the opposite lever, whichas shown below (figure 5) was required to drive performancebelow control levels. In addition, the effectiveness of thestimulation patterns was shown to decline across the same spanof extended delays in a manner similar to closed loop trialswith MIMO detected strong SR code firing patterns (figure 2).These results demonstrate that substitution of strong SR codepatterns of electrical stimulation eliminated trials that wereat risk for error when weak SR codes were detected by theMIMO model (figure 3).

3.3. MIMO model derived stimulation replaces lostmnemonic function

The above demonstration that MIMO model stimulation couldoverride the mismatch between anticipated duration of delayand strength of the SR code suggests that such stimulation,if synchronized and delivered at the time of occurrence ofthe SR, could provide a means of inducing strong SR codes.This was examined under the most severe condition by testinganimals when MIMO model predications from hippocampalCA3 inputs were not available online. As in previous studies(Hampson et al 1999a) normal operation of hippocampalcircuitry was seriously impaired by chronic infusion overa two week period of the glutamatergic transmissionblocking agent, MK801 (Collingridge et al 1983, Coanet al 1987). In animals that were trained to perform theDNMS task, infusion of MK801 unilaterally into the CA3region (37.5 μg h−1; 1.5 mg ml−1, 0.25 μl h−1) for 14 daysproduced significant disruptions in performance at alldelay intervals and significantly reduced the appearanceof strong SR codes detected online by the MIMO model(figure 3). However, since the MIMO model deliveredsuccessful SR stimulation patterns in the same animals undernormal (nondrug) circumstances, each animal’s previouslyidentified effective MIMO stimulation pattern was deliveredat the time the animal pressed the sample lever (SR)during testing while MK801 was being chronically infused.MK801 significantly suppressed DNMS performance at alldelays compared to control levels (F(1,416) = 13.37, p <

0.001). However, performance was significantly improved(F(1,416) = 9.52, p < 0.001) on trials in which CA1stimulation was delivered with strong SR code patterns, and

6

J. Neural Eng. 8 (2011) 046017 T W Berger et al

Figure 3. MIMO model controlled stimulation codes. Top: prediction of ‘weak’ SR encoding in CA1 via MIMO model results in CA1electrical stimulation with spatio-temporal patterns corresponding to strong SR code (green raster ‘substitution’, top right) predicted by priorclosed loop MIMO model applications (figure 2). CA1 stimulation consisted of biphasic electrical pulses (0.5-2.0 V, 10-50 μV, 0.5 sduration) delivered no more than once per 50 ms per channel, delivered on trials with delay durations of 15-50s. Bottom: comparison ofmean DNMS performance (% correct ±SEM) across animals (n = 23) on trials with interposed CA1 stimulation patterns derived fromMIMO generated strong SR codes (stim MIMO model) vs. trials on which no stimulation (no stim) was delivered. Trials in whichstimulation at the same intensity was generated from scrambled MIMO model CA1 coefficients (figure 2) are also shown.

remained only slightly (F(1,416) = 5.12, p = 0.02) belowcontrol levels. Figure 4 shows that delivery of MIMOstimulation patterns was highly effective in reversing thedetrimental effects of the drug and significantly increasedperformance on all trials with delays >10 s. Even thoughperformance was not elevated back to normal (pre-MK801)levels, delivery of effective MIMO model derived stimulationpatterns at the SR significantly reduced the drug-induceddeficit.

3.4. Demonstration of specificity of MIMO model derivedstimulation patterns

Together, the above demonstrations (figures 2–4) show thatdetection, imposition and closed loop control based on MIMO-derived ensemble firing patterns control DNMS task behavior

by enhancing encoding of the SR sufficient to counteract theeffects of the intervening delay interval. It is obvious thatsuch enhanced encoding facilitates making the NR decision,but the specificity of the MIMO model stimulation patterncan be assessed in the same experimental context. Adistinct test of the specificity of the information the SR codesand their strength was ascertained by reversing the MIMOstimulation patterns such that the pattern for the left leverwas delivered when the animal was presented with, and madethe SR, on the right lever and vice versa. Delivery of thestimulation pattern specific for encoding the lever oppositethe one presented and responded to sample, should not onlyeliminate enhanced encoding of the SR, but should actuallyimpair performance below normal levels due to stimulationinduced miscoding of information required in the nonmatchphase to correctly select the NR. Therefore the frequency of

7

J. Neural Eng. 8 (2011) 046017 T W Berger et al

Figure 4. MIMO model repair of hippocampal encoding with previously effective CA1 stimulation patterns delivered at the time of the SR.Top: intrahippocampal infusion (see figure 1, inset center right, cannula shown entering CA3 next to electrode array on right) ofglutamatergic antagonist MK801 for 14 days suppressed MIMO derived ensemble firing in CA1 and CA3 (red ‘X’, left). Strong SR codepatterns derived from previous DNMS trials were delivered bilaterally to the same CA1 locations when the SR occurred in the sample phase.Bottom: mean percentage correct (±SEM) performance (n = 5 animals) summed over all conditions, including vehicle-infusion (control)without stimulation, MK801 infusion (MK801) without stimulation, and MK801 infusion with the CA1 stimulation (substitute) pattern(MK801 + Stim).

errors (match responses) should increase in comparison totrials in which no stimulation occurred. Figure 5 shows thatby reversing the MIMO model SR stimulation patterns forthe lever presented in the sample phase, performance wasreduced below normal levels (F(1,731) = 12.53, p < 0.001)relative to control trials in which no stimulation occurred. Incontrast, trials within the same sessions in which SR codeswere delivered appropriate for the lever presented (normalstim) exhibited performance significantly above control levels(F(1,731) = 15.76, p < 0.001). The results shown infigure 5 support the conclusion that delivery of the opposite SRstimulation pattern was capable of overriding normal encodingtendencies by significantly impairing performance on trialswith deliberate mismatches between the SR stimulation patternand lever position. It is clear that reversed strong SR codepatterns had a large impact on performance. To validatethe fact that the MIMO stimulation patterns were effective,figure 5 also shows that performance was markedly improvedif the same SR stimulation patterns were delivered when theappropriate lever was presented in the sample phase. Onefinal control manipulation for specificity of the MIMO derived

stimulation patterns was to delay delivery of the stimulationby 3–5 s after occurrence of the SR (delayed stimulation,figure 5). The results show that if strong SR code stimulationpatterns were delayed by more that 3.0 s there was no effecton performance (stim late versus no stim: F(1,731) = 3.17,ns).

4. Discussion

The above demonstrations show that neural prosthesesemploying connections between different neuronal assembliescan be applied to accurately assess and impose effectivespatiotemporal task-related firing patterns which can be usedfor (1) closed loop control of performance (figure 2), (2)electrical stimulation of the same structures in the manner thatmimics output of successful task-related signals to facilitateand override instances in which less effective ensemble firingoccurs (figure 3), (3) recovery of function when the sameensembles are no longer operative (figure 4) and finally(4) validating whether the facilitative patterns of stimulationare specific to encoding of task-related events necessary for

8

J. Neural Eng. 8 (2011) 046017 T W Berger et al

Figure 5. Reverse stimulation to verify specificity of closed loop MIMO model stimulation. Top: validity of CA1 ‘strong SR code’ patternsdetermined by delivering stimulation patterns appropriate for the opposite type of DNMS trial (‘reversed’) or by varying the time at whichstimulation was delivered relative to the occurrence of the SR (‘delayed’). Bottom left: DNMS trials in which the SR code for the oppositelever was delivered (reversed stim) and compared with trials within the same sessions in which SR codes were delivered appropriate for thelever presented (normal stim) as well as control (no stimulation) trials. Bottom right: DNMS delay curves for animals stimulated withMIMO model generated strong SR codes either at the time of the SR (stim at SR) or commencing 3–5 s after the SR (stim late).

successful performance within specific contexts (figure 5).This type of utility provided by the MIMO model and appliedto a well-characterized set of ensemble firings provides oneof the first demonstrations in which deciphered nonlinearactivation patterns across cell assemblies can be utilized todeliver facilitative electrical stimulation to those same brainregions (Dulay et al 2009). Such an approach takes advantageof the fact that input–output relations are all that are requiredto make the predictions, irrespective of whether those relationsreflect mono- or polysynaptic connections. Therefore, even ifhippocampal CA3 connectivity to CA1 were not present, theMIMO model could be applied by characterizing input–outputrelations between dentate gyrus (and/or entorhinal cortex) andCA1 outputs if the former structures remained operative whenCA3 was damaged. In addition, by reversing the encodingpatterns delivered for critical behavioral events within thetask (figure 5), it was possible to test the specificity of theeffectiveness of the neural prosthesis for replacing damaged

or destroyed connections (Moritz et al 2008, Leuthardt et al2009), as shown in figure 4. While some factors remainto be determined as to the basis for the success of thisapproach (i.e. range of effective stimulation parameters, etc.),and even though opportunities for negative results wereprovided, the consistency of the outcomes of each of the abovetest conditions confirm the basic assumption of predictiveensemble encoding as performed by the MIMO model. Itwas not necessary to understand the basis for firing patternsin CA1 predicted by CA3 in terms of encoded information,even though in the past each structure has been implicatedindependently in performance of the task from many differentperspectives, including hippocampal removal, drug effectsand time course of behavioral acquisition (Hampson et al1999a, 2008, Deadwyler et al 2007). In addition, previousoffline assessments of data from the same DNMS taskemploying the exact same MIMO model (Zanos et al 2008,Song et al 2009) provide substantial computational support for

9

J. Neural Eng. 8 (2011) 046017 T W Berger et al

the online predictions of CA1 output firing patterns based onCA3 cell input firing to control online delivery of facilitativeSR stimulation patterns (figures 3 and 4).

The efficiency of the cortical prosthesis described hereutilizing a MIMO model that predicts firing patterns forsuccessful encoding of task specific information, provides aunique yet feasible approach to constructing prostheses toreplace information transmitted between brain structures vianonlinear patterned inputs and outputs (Song et al 2007a,2009). In the case of hippocampal based memories, the abovedemonstration shows that cognitive processes can be detectedin terms of a code that reflects the degree to which informationis successfully represented and encoded (figure 2) for retrievalat a later time (Wais et al 2006, Wixted 2007, Miller et al 2010).The fact that this code can be mimicked by delivery of electricalstimulation in the same spatio-temporal sequence during theto-be-remembered behavioral event, provides a means of (1)improving performance when encoding is deficient (figure 3),(2) replacing memory function when hippocampal processesare compromised (figure 4), and importantly (3) determiningthe specificity of the electrical stimulation used to mimic thefiring patterns that are not present (figures 3 and 5). What isalso apparent is that this cortical prosthesis does not requireseparate probes for delivery of effective neural stimulation,since the stimulation pulses were delivered through the sameCA1 array electrodes (figures 1 and 3). Therefore, it is likelythat similar assessments could yield further confirmation thatMIMO model derived stimulation patterns could be employedas a universal means of recovering lost connections betweenneural ensembles in other brain regions, as demonstrated hereutilizing previously well-characterized hippocampal firingpatterns.

Acknowledgments

The authors appreciate the efforts of Andrew Sweatt, VernellCollins, Rodrigo Espana, Li Hong Shi, Chad Collins andGeorge McLeod. This work was supported by DARPAcontracts to SD (N66601-09-C-2080) and to TB (N66601-09-C-2081) (Prog. Dir: COL Geoff Ling), and grants NSF EEC-0310723 to USC (TB), NIH/NIBIB grant no P41-EB001978to the Biomedical Simulations Resource at USC (VM and TB)and NIH R01DA07625 (SD).

References

Berger T W et al 2005 Restoring lost cognitivefunction—hippocampal-cortical neural prostheses IEEE Eng.Med. Biol. Mag. 24 30–44

Berger T W, Song D, Chan R H M and Marmarelis V Z 2010 Theneurobiological basis of cognition: identification bymulti-input, multi-output nonlinear dynamic modeling Proc.IEEE 98 356–74

Brown E N, Barbieri R, Venture V, Kass R E and Frank L M 2002The time-rescaling theorem and its application to neural spiketrain data analysis Neural Comput. 14 325–46

Coan E J, Saywood W and Collingridge G L 1987 MK-801 blocksNMDA receptor-mediated synaptic transmission and long termpotentiation in rat hippocampal slices Neurosci. Lett. 80 111–4

Collingridge G L, Kehl S J and McLennan H 1983 Excitatory aminoacids in synaptic transmission in the Schaffer

collateral-commissural pathway of the rat hippocampusJ. Physiol. 334 33–46

Daly J J and Wolpaw J R 2008 Brain-computer interfaces inneurological rehabilitation Lancet Neurol. 7 1032–43

Deadwyler S A, Goonawardena A V and Hampson R E 2007Short-term memory is modulated by the spontaneous release ofendocannabinoids: evidence from hippocampal populationcodes Behav. Pharmacol. 18 571–80

Deadwyler S A and Hampson R E 1995 Ensemble activity andbehavior: what’s the code? Science 270 1316–8

Deadwyler S A and Hampson R E 1997 The significance of neuralensemble codes during behavior and cognition Annu. Rev.Neurosci. 20 217–44

Deadwyler S A and Hampson R E 2004 Differential butcomplementary mnemonic functions of the hippocampus andsubiculum Neuron 42 465–76

Deadwyler S A and Hampson R E 2006 Temporal couplingbetween subicular and hippocampal neurons underliesretention of trial-specific events Behav. Brain Res.174 272–80

Dulay M F, Murphey D K, Sun P, David Y B, Maunsell J H,Beauchamp M S and Yoshor D 2009 Computer-controlledelectrical stimulation for quantitative mapping of humancortical function J. Neurosurg. 110 1300–3

Hampson R E and Deadwyler S A 2003 Temporal firingcharacteristics and the strategic role of subicular neurons inshort-term memory Hippocampus 13 529–41

Hampson R E, Jarrard L E and Deadwyler S A 1999a Effects ofibotenate hippocampal and extrahippocampal destruction ondelayed-match and -nonmatch-to-sample behavior in ratsJ. Neurosci. 19 1492–507

Hampson R E, Simeral J D, Berger T W, Song D, Chan R H Mand Deadwyler S A 2011 Cognitively relevant recording inhippocampus: beneficial feedback of ensemble codes in aclosed loop paradigm Electrophysiological RecordingTechniques ed R P Vertes and R W Stackman (New York:Humana Press) 215–39

Hampson R E, Simeral J D and Deadwyler S A 1999b Distributionof spatial and nonspatial information in dorsal hippocampusNature 402 610–4

Hampson R E, Simeral J D and Deadwyler S A 2008 Neuralpopulation recording in behaving animals: constituents of theneural code for behavior Neural Population Encodinged C Holscher and M H Munk (Cambridge: CambridgeUniversity Press) pp 74–94

Isa T, Fetz E E and Muller K R 2009 Recent advances inbrain–machine interfaces Neural Netw. 22 1201–2

Jarosiewicz B, Chase S M, Fraser G W, Velliste M, Kass R Eand Schwartz A B 2008 Functional network reorganizationduring learning in a brain–computer interface paradigm Proc.Natl Acad. Sci. USA 105 19486–91

Legenstein R, Chase S M, Schwartz A B and Maass W 2010 Areward-modulated hebbian learning rule can explainexperimentally observed network reorganization in a braincontrol task J. Neurosci. 30 8400–10

Leuthardt E C, Schalk G, Roland J, Rouse A and Moran D W 2009Evolution of brain–computer interfaces: going beyond classicmotor physiology Neurosurg. Focus 27 E4

Marmarelis V Z 2004 Nonlinear Dynamic Modeling ofPhysiological Systems (Hoboken, NJ: Wiley)

Marmarelis V Z and Berger T W 2005 General methodology fornonlinear modeling of neural systems with Poissonpoint-process inputs Math. Biosci. 196 1–13

Marmarelis V Z and Orme M E 1993 Modeling of neural systemsby use of neuronal modes IEEE Trans. Biomed. Eng.40 1149–58

Miller K J, Schalk G, Fetz E E, den Nijs M, Ojemann J G andRao R P 2010 Cortical activity during motor execution, motorimagery, and imagery-based online feedback Proc. Natl Acad.Sci. USA 2 4430–5

10

J. Neural Eng. 8 (2011) 046017 T W Berger et al

Moritz C T, Perlmutter S I and Fetz E E 2008 Direct control ofparalyzed muscles by cortical neurons Nature 456 639–42

Simeral J D, Hampson R E and Deadwyler S A 2005 Behaviorallyrelevant neural codes in hippocampal ensembles: detection onsingle trials Synaptic Plasticity: From Basic Mechanisms toClinical Applications ed M Baudry, X Bi and S Schreiber(Cambridge, MA: MIT Press) pp 459–76

Song D, Chan R H, Marmarelis V Z, Hampson R E, Deadwyler S Aand Berger T W 2007a Nonlinear dynamic modeling of spiketrain transformations for hippocampal-cortical prostheses IEEETrans. Biomed. Eng. 54 1053–66

Song D, Chan R H, Marmarelis V Z, Hampson R E,Deadwyler S A and Berger T W 2007b Statistical selection ofmultiple-input multiple-output nonlinear dynamic models ofspike train transformation Conf. Proc. IEEE Eng. Med. Biol.Soc. 1 4727–30

Song D, Chan R H, Marmarelis V Z, Hampson R E, Deadwyler S Aand Berger T W 2009 Nonlinear modeling of neural populationdynamics for hippocampal prostheses Neural Netw. 22 1340–51

Truccolo W, Eden U T, Fellows M R, Donoghue J P and Brown E N2005 A point process framework for relating neural spikingactivity to spiking history, neural ensemble, and extrinsiccovariate effects J. Neurophysiol. 93 1074–89

Velliste M, Perel S, Spalding M C, Whitford A S and Schwartz A B2008 Cortical control of a prosthetic arm for self-feedingNature 453 1098–101

Wais P E, Wixted J T, Hopkins R O and Squire L R 2006 Thehippocampus supports both the recollection and thefamiliarity components of recognition memory Neuron49 459–66

Wang W, Chan S S, Heldman D A and Moran D W 2010a Motorcortical representation of hand translation and rotation duringreaching J. Neurosci. 30 958–62

Wang W, Collinger J L, Perez M A, Tyler-Kabara E C, Cohen L G,Birbaumer N, Brose S W, Schwartz A B, Boninger M Land Weber D J 2010b Neural interface technology forrehabilitation: exploiting and promoting neuroplasticity Phys.Med. Rehabil. Clin. N. Am. 21 157–78

Witter M P and Amaral D G 2004 The Rat Nervous System 3rd edned G Paxinos (San Diego, CA: Academic) pp 637–87

Wixted J T 2007 Dual-process theory and signal-detection theory ofrecognition memory Psychol. Rev. 114 152–76

Zanos T P, Courellis S H, Berger T W, Hampson R E,Deadwyler S A and Marmarelis V Z 2008 Nonlinear modelingof causal interrelationships in neuronal ensembles IEEE Trans.Neural Syst. Rehabil. Eng. 16 336–52

11